An Adaptive Two-level Filtering Technique for Noise Lines in Video Images

Baris Baykant ALAGOZ & Mehmet Emin TAGLUK
International Journal of Image Processing, (IJIP) Volume (5) : Issue (3) : 2011 270
An Adaptive Two-level Filtering Technique for Noise Lines in
Video Images
Baris Baykant ALAGOZ baykant.alagoz@inonu.edu.tr
Department of Electrical and Electronics Engineering,
Inonu University, Malatya, Turkey
Mehmet Emin TAGLUK mehmet.tagluk@inonu.edu.tr
Department of Electrical and Electronics Engineering,
Inonu University, Malatya, Turkey
Abstract
Due to narrow-band noise signals in transmission channels, visible lines of disturbance can
appear in video images. In this paper, an adaptive method based on two-level filtering is
proposed to enhance the visual quality of such images. In the first level, an adaptive orientation
selective filter detects and clears the noisy lines in the image. In the second level, a median filter
repairs defects resulting from the orientation selective filtering process and also filters the wide-
band impulsive noise. It was observed that in case of periodic noisy lines in TV images, this
filtering technique can sufficiently enhance the image quality and improve the SNR level.
Keywords: Adaptive noise filter, Wireless video image enhancement.
1. INTRODUCTION
In long distance wireless video transmission systems, periodic noise line patterns commonly
appear on the received image. Such noise line patterns in wireless transmission are mostly
caused from long-term narrow-band signal interference to the communication channel. In many
cases, such interference of by noise signals in the channel is unavoidable and therefore the
removal of these noise signals has to be carried out on the received image by using filtering
techniques. A basic notch reject filter has been applied for the removal of the noise lines on the
images received from the long-distance space missions [1]. Nowadays, with increasing usage of
wireless video transmission systems in day-to-day applications there is an increasing demand to
develop such filtering techniques to work on received images.
In practice, unlicensed transmitter interference, noise in electronics, multi-path effects, loss of
horizontal or vertical synchronization are all seen to cause periodic noise lines on the received
image. These noise lines can severely mislead computer vision algorithms employed in
autonomous remote control systems used in unmanned vehicles [2-5].
Periodic noise lines in images are commonly seen in imaging systems, which use a row scanning
mechanism in the construction of the image data when a long-term noise signal affects the
system. For example, mechanical and acoustic vibrations in force sensors were seen to decrease
the signal to noise ratio (SNR) of images scanned by an Atomic Force Microscope (AFM) at video
rate [6].
In the image enhancement field, a variety of methods have been developed to filter out the effects
of random noise on images. Most of these studies have been focused on preserving singular
features of the image such as edges, while smoothing other segments of the image [7-11].
Alternatively a nonlinear adaptive filter based on a neural network has been proposed for
reducing the additive noise [12]. Many adaptive image restoration methods that analyse the noise
and optimise the behaviour of the filter to improve overall filtering performance [13-15].
Specifically, for digital TVs the filters compromise edge detection, and an automated modification
of filter coefficients has been addressed in detail by Chan et al. [7, 10]. These filters were mainly

developed to deal with random additive noise in images. In their mathematical formulation, the
received signal has been defined as:
)()()( xnxuxur += , (1)
where, )(xn is the random noise and )(xu is the original noise-free image. Due to the fact that
the noise components are treated as a part of the image signal )(xu , such filters designed for
removing random noise are not able to effectively deal with narrow-band noise lines.
In our study, we assume that the received signal has a narrow-band noise signal besides
a random noise signal. Under this assumption, a noise signal can be modelled as:
)()()()( xnxnxuxu nr ++= (2)
where, )(xnn represents the narrow-band noise.
To enhance the received image coming from highly noisy channel modelled by equation (2), we
proposed a filtering structure which is composed of a frequency domain orientation selective
notch reject filter for the elimination of the narrow-band noise ( )(xnn ) [1, 16] and a spatial
median filter for impulsive noise [8]. The block diagram of such a two-level filtering is presented in
Figure 1.
FIGURE 1: Block diagram of two-level filtering.
A noisy image is first transformed to red, green and blue color channels, each color channels
being solely filtered by the two-level filter designed. In order to adapt the angle of notch reject
band to the slope of the noise lines in the received images, the proposed orientation selective
filter first detects the spectral region belonging to noise lines, via measuring the local spectral
power density, and then it orients the reject band onto the region where the power of noise lines
is intensified.
2. PROPOSED METHOD
Frequency domain filtering [16] usually provides a good performance in the filtering of narrow-
band noise signals, because of allow us eliminating frequency components in a narrow frequency
band. Basic design methodology of a frequency domain filter is as follows: The spatial two-
dimensional image data are first transformed into frequency components via a two-dimensional
Fast Fourier Transform (2DFFT) and then a two-dimensional mask is applied to suppress
undesired frequency components, and finally an inverse Fast Fourier Transform (2DIFFT) is used
to obtain the filtered image in the spatial domain, as illustrated in Figure 2.
Frequency
Domain
Filter
u(x)+nn(x)+n(x)
Two-level filtering system
Spatial
Filter
u(x)+n(x) uo(x)

FIGURE 2. Frequency domain filtering with component masking
Masking is done by multiplication of each mask element with elements of R , G and B image
as follows:
),(),(),( vuMkvuRvuR ffm ⋅= ,
),(),(),( vuMkvuGvuG ffm ⋅= , (3)
),(),(),( vuMkvuBvuB ffm ⋅= ,
where, fR , fG and fB are the FFT of the R , G and B matrix, fmR and , fmG and fmB
are the masked frequency components of the RGB image. The mask matrix Mk has real value,
RvuMk ∈),( , in such cases this mask takes effect on the amplitude of the spectral components
of images. The phases of frequency components are preserved in this masking operation.
In this study, we introduce a mask φMk generation function that provides a directional stop-
band over high frequencies to suppress periodic line patterns in the image for a given φ angle.
The proposed mask is formed from superposition a suppressing channel passing through the
origin of the 2D spectrum and a directional pass band at the origin.
The construction of the mask is as follows. The suppressing channel, whose direction is
controlled by a given angle, φ , was constructed on a base line passing through the origin
)
2
,
2
(
NM
as presented in Figure 3.
2
)
2
()tan(
MN
vu +−⋅−= φ , (4)
where M and N determine the size of the mask, φMk , and φ is angle of the base line. The
shortest distance of any point ),( vu to this base line, denoted by ),,( φvuds , was derived as:
2-Dimensional
Fast Fourier
Transform
(FFT)Digital
Đmage
Component
Masking
Frequency Domain Filtering
Filtered
Đmage
2-Dimensional
Inverse Fast
Fourier
Transform
(IFFT)










=−
=−
<<−+−
=
0
0
0022
270,
2
90,
2
27090,)),,(()),,((
),,(
φ
φ
φφφ
φ
M
u
N
v
vuvvvuuu
vud
bb
s , (5)
where )),,(),,,(( φφ vuvvuu bb is the nearest point of the base line to the point ),( vu on the
mask, and calculated by the following expressions:
22
),,()tan(),,(
MN
vuvvuu bb +





−⋅= φφφ . (6)
)1)((tan2
)tan(2)tan(2)(tan
),,( 2
2
+⋅
⋅⋅+⋅−⋅+⋅
=
φ
φφφ
φ
uMvN
vuvb . (7)
The shortest distance of any point ),( vu to the centre of the mask, denoted by ),( vudc , is
calculated as:
22
)
2
()
2
(),(
N
v
M
uvudc −+−= . (8)
The mask generation function is then formed as:
)
),(
exp()
),,(
exp())
),,(
exp(1(),( 2
2
2
2
2
s
s
c
c
s
s vudvudvud
vuMk
αα
φ
α
φ
φ −⋅−+−−= , (9)
or in a more compact form:
)
),,(
exp(1)
),,(
exp(1),( 2
2
2
2
s
s
c
c vudvud
vuMk
α
φ
α
φ
φ −⋅







−−+= , (10)
where sα and cα are the standard deviations determining the width of directional suppression
channel. These parameters also control smoothness of the transition from the stop to pass band
of the filter. This smooth transition consequently eliminates ringing effects appearing around
patterns in the image. In Figure 4, Mk matrix generated for various φ are illustrated.
The angle φ is adaptively determined in a range of [ ]maxmin ,φφ by applying the following steps:
Step 1: Calculate )(φPw for all φ angles from minφ to maxφ . ( )(φPw , expressed by equation
(11), is the average spectral power in the image spectrum that is effected by the suppression
channel of φMk )
Step 2: Use the φMk mask in filtering for the φ , at which )(φPw is a maximum and value of this
maximum exceeds a predefined filter activation threshold ( thP ). In the case, all )(φPw
[ ]maxmin ,φφφ ∈ are lower than thP , the current image can bypass filtering. This implies that the
noise lines do not have enough power to activate filtering.

FIGURE 3. A contour plot of a suppression channel and its relevant parameters with respect to a base line
passing through the centre of the spectrum.
The average image power in the suppression channel )(φPw can be calculated by:
∑∑
∑∑
−
=
−
=
−
=
−
=
++⋅−
−
=
1
0
1
0
222
1
0
1
0
)),(),(),(()),,(1(
)),,(1(
1
)(
M
u
N
v
fffM
u
N
v
vuBvuGvuRvuMk
vuMk
Pw φ
φ
φ (11)
The term )),,(( φvuMk−1 in equation (11) is the weighting function for spectral components.
Equation (11) is a discrete two dimensional extension of a single variable weighted average
spectral power given as:
dffIfw
dffw
Pw ∫
∫
⋅=
2
)()(
)(
1
for a weight function )( fw .
The choice of cα and sα parameters are in fact based on trial and error. However it was
observed that, in this particular application, for TV images with a size of 640x480 pixel, then
2828=cα and 316=sα can give a satisfactory results.
After removing the noise line patterns from the image by adaptive frequency domain filtering, a
3x3 median filter is applied to the image in order to improve the image quality. This median
filtering reduces the deformations resulting from the removal of noise line patterns, as well as the
impulsive noise on the filtered image.

FIGURE 4. (a) Mk matrix generated for
0
90=φ (b) Mk matrix generated for
0
120=φ and (c)
Mk matrix generated for
0
80=φ
3. EXPERIMENTAL RESULTS
In Figure 5 and in Figure 6, TV images from a real media broadcast captured by a commercial TV
Card were enhanced by the adaptive two-level filtering process described in the previous
sections. The calculated spectral power as a function angle φ and the resulting Mk mask are
given in the sub-figures (c) and (d), respectively.
(a)
100 200 300 400
50
100
150
200
250
300
350
400
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Allowed
Suppressed
Φ
ΦminΦmax
(c)(b)
100 200 300 400
100
200
300
400
100 200 300 400
100
200
300
400

FIGURE 5. (a) Noise test image from TV, (b) Enhanced image by filter system ( 2828=cα , 316=sα
) , (c) Generated Mk matrix, and (d) )(φPw values in the adaptation process. The filter applied a mask
for φ =
0
90 .
(a)
(b)
100 200 300 400 500 600
100
200
300
400
500
Mk
(c)
70 80 90 100 110
0
0.5
1
1.5
2
2.5
x 10
10
Φ
Pw(Φ)
Pth
[ Decree ]
(d)

FIGURE 6. (a) Noise test image from TV, (b) Enhanced image by filter system ( 2828=cα , 316=sα
) , (c) Generated Mk matrix, and (d) )(φPw values in the adaptation process. The filter applied a mask
for φ =
0
90 .
(a)
100 200 300 400 500 600
100
200
300
400
500
Mk
(c)
70 80 90 100 110
0
5
10
15
x 10
9
Φ
Pw(Φ) Pth
[ Decree ]
(d)
(b)

FIGURE 7. (a) and (b) are noise test images and their frequency spectrums , (c) and (d) are adaptive
orientation selective filtering results and their frequency spectrums , (e) and (f) are median filtered images of
(c) and their frequency spectrums..
In Figure 7, the effects of filtering at each level on the test image and its frequency spectrum are
illustrated. After cleaning noise line patterns from the image, as seen in Figure 7(c), median
filtering with 3x3 windowing can be seen to effectively restore the image from impulsive noise and
deformation resulting from the frequency filtering. Median value with respect to image data in 3x3
windows have an effect of repairing notch filtering that is a vertical suppression band passing
through the centre in Figure 7(d). This effect can be seen as strengthening some of the frequency
components in the vertical suppression band in Figure 7(f).
4. TEST ON SYNTHETIC IMAGE AND SNR LEVELS
In this section, the noisy television test image shown in Figure 8(a) was distorted by a synthetic
noise line pattern. This additive noise was generated by a sinusoidal signal with an amplitude of
80 and a frequency set to 0.001, as in Figure 8(c), and a pseudo-random noise signal with an
uniform distribution in the interval ( )50,50− , as in Figure 8(d). After adding these noise signals
to the television test image in accordance with Equation 2, a noise test image was obtained as
shown in Figure 8(b).

FIGURE 8. (a) and (b) are noise test images and their frequency spectrums , (c) and (d) are adaptive
orientation selective filtering results and their frequency spectrums , (e) and (f) are median filtered images of
(c) and their frequency spectrums.
The noisy television test image given in Figure 8(b) was enhanced by using three different basic
filters and the proposed two-level filtering, and the results obtained are demonstrated in Figure 9.
In Figure 9 it can be seen that the median-filter and Wiener filter with window size of 3x3 did not
sufficiently remove the line patterns in the test image. Even though the Gaussian low pass filter
blurred the image, the noise lines are still visible. The proposed method detected the noise lines
at
0
83=φ and adapted the orientation selective filter according to the noise lines. After the
application of both orientation selective filter in the frequency domain and the median filter in the
spatial domain the noise lines were removed without too much blurring of the image as seen in
Figure 9(d).

Figure 9. (a).Median filtered image, (b) Wiener filtered image, (c) Low pass filtered image by a Gaussian
function mask, (d) The proposed two-level filtered image.
(d)
Mk
50 100 150 200 250 300
50
100
150
200
70 80 90 100 110
0
2
4
6
8
10
x 10
9
Pth
Φ
Pw(Φ)
[ Decree ]

FIGURE 10. Generated Mk matrix and )(φPw values in the adaptation process. The filter
was found to be best when φ =
0
83 .
In Figure 11, we present Signal to Noise Ratio (SNR) of these filters on the noisy television test
image. In these tests, the thP of the orientation selective filter activation was set to
9
10.3 . In this
case, when the SNR of the received image is over 40 dB, the noise level remains below thP and
the orientation selective filtering was bypassed and the adaptive two level filtering performed the
median filtering on the received image, solely. When the SNR of the test image was decreased to
lower levels, the proposed two-level filtering detected the visible noise line patterns in the
received image and removed them from the image. In this way, it maintains higher SNRs in the
case of a video transmission in highly noisy channel.
FIGURE 11. SNR of the filtered images for various additive noise levels.
5. CONCLUSIONS
The noise channel model given by equation (2) provides a more realistic model of today’s
communication channels. In this manner, a more adaptive and sophisticated filtering system is
needed to restore video images. The proposed adaptive two-level filtering method, which
integrates an adaptive orientation selective filter and a median filter, was seen to effectively
enhance video images containing visible periodic noise lines.
The results obtained show us that the narrow-band noise signal can be detected and filtered by
mask adaptive frequency domain filtering and the slight deformations resulting from such
frequency domain filtering on the image data can be repaired by the median filtering.
6. REFERENCES
[1] R.C. Gonzalez, R.E. Woods, S.L. Eddins. Digital image processing using MATLAB. New
Jersey:Prentice Hall, 2004.
[2] J. Byrne, R. Mehra.” Wireless video noise classification for micro air vehicles.” Proceedings
of the 2008 Association for Unmanned Vehicle Systems International (AUVSI) Conference,
2008, pp.1-25.
10 15 20 25 30 35 40 45
15
20
25
30
35
40
45
SNR of Received Image
SNRofFilteredImage
Adaptive two-level Filtered Image
Wiener Filtered Image
Median Filtered Image
Gauss Filtered Image

[3] P.Y. Oh, W.E. Green, G. Barrows. “Flying insect inspired vision for autonomous aerial robot
maneuvers in near-earth environments.” IEEE International Conference on Robotics and
Automation (ICRA) , 2004, pp.2347–2352.
[4] J. Byrne, M. Cosgrove, R. Mehra. “Stereo based obstacle detection for an unmanned air
vehicle.” IEEE International Conference on Robotics and Automation (ICRA’06), 2006.
[5] J. Byrne, C.J. Taylor. ”Expansion segmentation for visual collision detection and Estimation.”
IEEE International Conference on Robotics and Automation Kobe International Conference
Center, 2009, pp.875-882.
[6] G. Schitter, M.J. Rost. ”Scanning probe microscopy at video-rate.” Materials Today, vol.11,
pp.40–48, 2008.
[7] T.F. Chan, S. Osher, J. Shen. “The digital TV filter and nonlinear denoising.” Image
Processing IEEE Transaction, vol.10, pp.231-241, 2001.
[8] C.A. Glasbey, G.W. Horgan. Image analysis for the biological science. Statistics nn practice,
New York:John Wiley & Sons, 1995.
[9] B. Jaehne. Digital image processing, Concepts, algorithms, and scientific applications,
Berlin:Springer, 1997.
[10] S. Osher, J. Shen. Digitized PDE method for data restoration, In: G. A. Anastassiou (Ed.),
Analytic-computational methods in applied mathematics, Chapman & Hall/CRC, 2000.
[11] A. Lev, S.W. Zucker, A Rosenfelt. ”Iterative enhancement of noisy images.” IEEE
Transanctions on Systems, Man and Cybernetichs, vol.7, pp.435-443, 1977.
[12] L. Corbalan, G.O. Massa, C. Russo, L. Lanzarini, A. De Giusti. “Image recovery using a
new nonlinear adaptive filter based on neural networks.” Journal of Computing and
Information Technology CIT, vol.14, pp.315–320, 2006.
[13] H.S. Wong, L. Guan. “A neural learning approach for adaptive image restoration using a
fuzzy model-based network architecture.” Neural Networks IEEE Transactions, vol.12,
pp.516–531, 2001.
[14] E.S. Hore, B. Qiu, H.R. Wu. “Adaptive noise detection for image restoration with a multiple
window configuration.” Image Processing Proceedings. 2002 International Conference,
2002, pp. 329-332.
[15] F. Russo. “A method for estimation and filtering of Gaussian noise in images.”
Instrumentation and Measurement, IEEE Transactions, vol.52, pp.1148–1154, 2003.
[16] R.P. Matei. “Design method for orientation-selective CNN filters.” Circuits and Systems,
Proceedings of the 2004 International Symposium vol.3, 2004, pp.105-8.

An Adaptive Two-level Filtering Technique for Noise Lines in Video Images

More Related Content

What's hot (20)

Viewers also liked (12)

Similar to An Adaptive Two-level Filtering Technique for Noise Lines in Video Images (20)

Recently uploaded (20)

An Adaptive Two-level Filtering Technique for Noise Lines in Video Images